CoolFace
Modelpublic

amwangfan/privacy-gateway-v4-qwen2.5-0.5b

sourceHugging Faceapache-2.0updated 4d agoView on Hugging Face
0likes153downloads
Model Card

Qwen2.5-0.5B-Privacy-Gateway-v4 (step-220)

简体中文 | English

网关： https://github.com/amwangfan/privacy-gateway

本地隐私网关的残差 SECRET / SAFE 分类器。三层分工：Layer 0 只抓形态极稳的 sk- / PEM / JWT / ghp_ / AKIA / 带口令的 URI；password= 只是把右边的值送进模型，不是硬替换；本 LoRA 给漏网 span 打分。弱口令 Layer 1 可以判 SAFE（漏了就漏了），不要打开「凡是 password= 就替换」。

  • —只训 attention / MLP（不含 `lm_head`），r=16，α=32
  • —检查点：step-220（v6.2 再平衡），合并为 N100 用的 F16 / Q8 GGUF
  • —训练与线上同一套短 prompt：
Secret? k={key} v={span} c={ctx} ->
Secret? v={span} c={ctx} ->
  • —下一 token：SECRET=65310，SAFE=83788
  • —空 c= → SAFE。生产环境：有键用 c=env file，无键残差用 c=bearer token。

启用探针（只卡假阳性）

网关 auto 模式仅当所有 SAFE 用例仍判 SAFE 才启用本模型（ok = negatives_kept == n_neg）。

  • —名字、路径、备份名、版本号、session id 不得被标成 SECRET。
  • —弱口令（如 mysql_root_password_2026）Layer 0 拦不住（对不上硬正则）；Layer 1 判 SAFE 就会出网。这是预期，不是 bug。
  • —Layer 0 已覆盖的形态（sk-…、JWT eyJ…）不要求 Layer 1 抓全。

N100 实测（F16，2026-09-18）：negatives_kept=12/12，positives_caught=2/6，decision=model。

文件

文件用途
privacy-gateway-v4-lora-step220.zipPEFT adapter（step-220）
adapter_config.json / adapter_model.safetensorsPeftModel.from_pretrained
qwen2.5-0.5b-privacy-v4-f16.ggufN100 生产（llama-server）
qwen2.5-0.5b-privacy-v4-q8_0.gguf更小的备用量化

本仓库不含训练 jsonl 与真实凭据。

相关链接

Apache-2.0

<a id="english"></a>

Qwen2.5-0.5B-Privacy-Gateway-v4 (step-220)

简体中文 | English

Gateway: https://github.com/amwangfan/privacy-gateway

Residual SECRET / SAFE classifier for a local privacy gateway. Three layers: Layer 0 only matches stable shapes (sk- / PEM / JWT / ghp_ / AKIA / password-in-URI). password= forwards the RHS to this model; it is not a hard replace. Weak passwords may be SAFE here (a miss is a miss). Do not enable “replace everything after password=”.

  • —LoRA on attention/MLP only (no `lm_head`), r=16, α=32
  • —Checkpoint: step-220 (v6.2 rebalance), merged to F16/Q8 GGUF for N100
  • —Train == serve prompt:
Secret? k={key} v={span} c={ctx} ->
Secret? v={span} c={ctx} ->
  • —Next-token ids: SECRET=65310, SAFE=83788
  • —Empty c= → SAFE. Production uses c=env file (keyed) or c=bearer token (keyless residual).

Readiness probe (FPR only)

The gateway auto mode enables this model iff every SAFE probe case stays SAFE (ok = negatives_kept == n_neg).

  • —Names, paths, backups, versions, session ids must not be flagged.
  • —Weak passwords (e.g. mysql_root_password_2026) are not Layer 0; if Layer 1 says SAFE they leave the host. That is expected.
  • —Layer-0-shaped tokens (sk-…, JWT eyJ…) are not required of Layer 1.

Live N100 (F16, 2026-09-18): negatives_kept=12/12, positives_caught=2/6, decision=model.

Files

fileuse
privacy-gateway-v4-lora-step220.zipPEFT adapter (step-220)
adapter_config.json / adapter_model.safetensorsPeftModel.from_pretrained
qwen2.5-0.5b-privacy-v4-f16.ggufN100 production (llama-server)
qwen2.5-0.5b-privacy-v4-q8_0.ggufsmaller fallback

Training jsonl and real credentials are not in this repo.

Related

Apache-2.0